Summary

Replace three Python str-method calls in Position.gen_moves with literal-string in checks. The search tree is unchanged; the only thing that changes is how each char is classified inside the move-generator loop.

Why

gen_moves accounted for ~67% of CPython search time in a cProfile run of a 5-ply search from startpos. Inside it, the three str-methods above are called millions of times per search and each one is a Python-level attribute lookup plus a C call. Literal in is faster because CPython has a specialized opcode for it (CONTAINS_OP) that skips the method-lookup overhead.

timeit over a 120-char board scan, 100k iters, CPython 3.13:

c.isupper()                  3.33 us
c in "PNBRQK"                2.61 us   (~22% faster)
c.isspace() or c.isupper()   4.97 us
c in " \nPNBRQK"             3.34 us   (~33% faster)
c.islower()                  ~     (similar order)
c in "pnbrqk"                ~

Speedup

A 6-position suite (startpos + 5 typical openings/middlegames) at fixed depth 5, 5 runs each, CPython 3.13 on Windows:

              original          patched
nps           16,711 – 17,240   17,592 – 17,685
mean          17,030            17,656
speedup       —                 +3.7%

The variance bands don't overlap across 5 runs, so this is reproducible — not noise.

Correctness

• Node-count identical: every position at every depth produces the same node count between original and patched (123,142 across the depth-5 suite, both versions).
• Perft: matches across 6 standard positions at depth 3 — startpos = 8902, Kiwipete = 97862, position 3 = 2812, position 4 = 9467, promotion test = 62379, middlegame = 89890. Per-root-move breakdowns are also identical.
• Mate-in-1 puzzles: first 8 entries from tools/test_files/mate1.fen produce identical bestmove between original and patched.

Notes on minimalism

build/clean.sh sunfish.py | wc -l is unchanged: 3 lines swapped 1:1 in gen_moves. The diff also adds a 3-line # NB: comment explaining why the literal form is used (so a future reader doesn't "clean up" back to method calls); comments are stripped by clean.sh so the 131-line claim is unaffected.

Test plan

• Run perft_ab.py (modified vs original) at depth 3 across 6 positions — totals and per-root-move breakdowns identical.
• Run mate-in-1 cross-check on first 8 mate1.fen positions — identical bestmove.
• Run bench.py at depth 4 and depth 5, 5 runs each — speedup is reproducible.
• (Upstream maintainer can verify) tools/quick_tests.sh end-to-end.